JP7454402B2 - Detection system and learning method - Google Patents

Description

The present invention relates to a detection system including a Cherenkov detector and a signal processing unit using a neural network, and a method for training the neural network.

When the moving speed of charged particles (eg, electrons, muons) in a radiator is faster than the speed of light in the radiator, light is emitted due to the interaction between the charged particles and the radiator. In addition, uncharged particles (e.g., gamma rays, neutrinos) may generate charged particles such as electrons due to the photoelectric effect in a radiator, and the moving speed of the charged particles is faster than the speed of light in the radiator. If it is fast, light is emitted. Such a phenomenon is called Cherenkov effect or Cherenkov radiation, and the emitted light is called Cherenkov radiation or Cherenkov light.

A Cerenkov detector can detect arriving particles by using the Cerenkov effect. A Cerenkov detector includes a radiator that generates Cerenkov light through interaction with incident particles, and a photodetector that detects the Cerenkov light generated by the radiator. A Cerenkov detector is used, for example, in a PET (Positron Emission Tomography) device or a SPECT (Single Photon Emission Computed Tomography) device as a radiation detector that detects γ rays (energy 511 keV in PET) arriving from a subject.

Non-Patent Document 1 describes a technique for estimating an interaction position in a radiator using a neural network based on a Cerenkov light detection event by a photodetector of a Cerenkov detector.

Hashimoto, F., et al. “Afeasibility study on 3D interaction position estimation using deep neuralnetwork in Cherenkov-based detector: a Monte Carlo simulation study.”Biomedical Physics & Engineering Express, 5 (2019): 035001.

In the above technique, it is necessary to train a neural network using a large amount of training data, and then estimate an interaction position using the trained neural network. However, since it takes a long time to collect a large amount of training data used for learning, learning a neural network is not easy.

The present invention has been made to solve the above problems, and provides an apparatus and method for easily learning a neural network in a detection system equipped with a Cerenkov detector and a signal processing unit using a neural network. The purpose is to provide

The detection system according to the first aspect of the present invention includes: (1) a radiator that generates Cerenkov light by interaction with incident particles; and a light receiving surface that detects the Cerenkov light generated by the radiator; (2) a photodetector that outputs a signal representing the detection position and detection time of Cerenkov light; , an estimation unit that estimates the interaction position on the radiator using a neural network based on the number of Cerenkov light detection events on the light receiving surface and the detection position and detection time of each Cherenkov light detection event, and (3) for each interaction event. A plurality of sets of less than n Cherenkov light detection events are generated from among the n Cherenkov light detection events by the photodetector of the Cherenkov detector, and each Cherenkov light detection event included in each of the plurality of sets is generated. (4) a learning unit that causes a neural network to learn using the learning data created by the learning data creating unit; and (4) a learning unit that uses the learning data created by the learning data creating unit.

A detection system according to a second aspect of the present invention includes: (1) a radiator that generates Cerenkov light by interaction with incident particles; and a light receiving surface that detects the Cerenkov light generated by the radiator; (2) a first Cherenkov detector and a second Cherenkov detector, each including a photodetector that outputs a signal representing the detection position and detection time of Cherenkov light; The signal output from the photodetector is input, and for each coincidence event of photon pairs generated by the annihilation of a positron and an electron, the (3) a first Cerenkov detector for each coincidence event; (3) a first Cerenkov detector for each coincidence event; Generate a plurality of sets of less than n1 Cherenkov light detection events from among the n1 Cherenkov light detection events by the photodetector, and determine the detection position and detection of each Cherenkov light detection event included in each of the plurality of sets. Using the time as the first learning data, a plurality of sets of Cerenkov light detection events of less than n2 are generated from among the n2 Cerenkov light detection events by the photodetector of the second Cherenkov detector, and these multiple sets are (4) a learning data creation unit that uses the detection position and detection time of each Cherenkov light detection event contained therein as second learning data; and (4) the first learning data and the second learning data created by the learning data creation unit. A learning unit that causes the neural network to learn using learning data.

The learning method of the first aspect of the present invention includes (1) a radiator that generates Cerenkov light by interaction with incident particles, and a light-receiving surface that detects the Cerenkov light generated by the radiator; (2) a photodetector that outputs a signal representing the detection position and detection time of Cerenkov light; a neural network for a detection system, comprising: an estimator that estimates an interaction position on a radiator using a neural network based on the number of Cerenkov light detection events on a light receiving surface and the detection position and detection time of each Cherenkov light detection event; This is a method for making children learn. The learning method of the first aspect of the present invention includes (a) for each interaction event, a plurality of sets of less than n Cerenkov light detection events from among n Cerenkov light detection events by the photodetector of the Cherenkov detector; (b) learning data created in the learning data creation step; and a learning step of training the neural network using the data.

The learning method of the second aspect of the present invention includes (1) a radiator that generates Cerenkov light by interaction with incident particles, and a light-receiving surface that detects the Cerenkov light generated by the radiator; (2) a first Cherenkov detector and a second Cherenkov detector, each including a photodetector that outputs a signal representing the detection position and detection time of Cherenkov light; By inputting the signal output from the photodetector of The present invention is a method for training a neural network of a detection system including an estimator that estimates a photon pair generation position using a neural network based on the number of photodetection events and the detection position and detection time of each Cherenkov photodetection event. The learning method of the second aspect of the present invention includes: (a) for each coincidence event, a set of less than n1 Cherenkov light detection events from among n1 Cherenkov light detection events by the photodetector of the first Cherenkov detector; The detection position and detection time of each Cerenkov light detection event included in each of these multiple sets is used as first learning data, and n2 Cerenkov light detections are performed by the photodetector of the second Cerenkov detector. A learning method in which a plurality of sets of less than n2 Cerenkov light detection events are generated from among the events, and the detection position and detection time of each Cerenkov light detection event included in each of the plurality of sets is used as second learning data. The method includes a data creation step, and (b) a learning step of learning a neural network using the first learning data and second learning data created in the learning data creation step.

According to the present invention, in a detection system including a Cerenkov detector and a signal processing section using a neural network, the neural network can be easily trained.

FIG. 1 is a diagram showing the configuration of a detection system 1 according to the first embodiment. FIG. 2 is a perspective view showing the configuration of the Cherenkov detector 3. FIG. 3 is a diagram illustrating an example of processing by the estimation unit 42 of the signal processing unit 4 when estimating the interaction position. FIG. 4 is a diagram illustrating an example of processing by the learning data creation unit 43 and learning unit 44 of the signal processing unit 4 during learning of the neural network. FIG. 5 is a diagram showing an example of the configuration of a neural network. FIG. 6 is a diagram showing the results of the simulation. FIG. 7 is a diagram showing the configuration of the detection system 2 of the second embodiment. FIG. 8 is a diagram illustrating an example of processing performed by the estimation unit 52 of the signal processing unit 5 when estimating the photon pair generation position. FIG. 9 is a diagram illustrating an example of processing by the learning data creation unit 53 and learning unit 54 of the signal processing unit 5 during learning of the neural network. FIG. 10 is a diagram showing an example of the configuration of a neural network. FIG. 11 is a diagram showing the results of the simulation.

DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments for carrying out the present invention will be described in detail with reference to the accompanying drawings. In addition, in the description of the drawings, the same elements are given the same reference numerals, and redundant description will be omitted. The present invention is not limited to these examples, but is indicated by the claims, and is intended to include all changes within the meaning and scope equivalent to the claims.

FIG. 1 is a diagram showing the configuration of a detection system 1 according to the first embodiment. The detection system 1 includes a Cerenkov detector 3 and a signal processing section 4. Cherenkov detector 3 includes a radiator 10 and a photodetector 20. FIG. 2 is a perspective view showing the configuration of the Cherenkov detector 3.

The radiator 10 generates Cerenkov light through interaction with incident particles (eg, γ rays). The radiator 10 is made of a monolithic material with a known refractive index and capable of producing the Cerenkov effect. The radiator 10 is preferably made of a material that does not easily generate scintillation light. The material of the radiator 10 is, for example, lead glass (SiO ₂ +PbO), lead fluoride (PbF ₂ ), PWO (PbWO ₄ ), or the like. Preferably, the radiator 10 has a rectangular parallelepiped shape.

The photodetector 20 has a light receiving surface 21 that detects Cerenkov light generated by the radiator 10. The light-receiving surface 21 of the photodetector 20 faces one surface of the rectangular parallelepiped-shaped radiator 10 and is parallel to that surface. The light receiving surface 21 of the photodetector 20 may be in contact with one surface of the radiator 10. Based on the detection of Cherenkov light on the light receiving surface 21, the photodetector 20 outputs a signal representing the detection position (x, y) of the Cherenkov light on the light receiving surface 21 and the detection time t. The signal can represent the detection time t by the pulse output at the time of Cerenkov light detection.

Preferably, the photodetector 20 is a high-speed, high-sensitivity device capable of photon counting operation. As the photodetector 20, for example, a multi-anode photomultiplier tube or MPPC (registered trademark) is preferably used. A multi-anode photomultiplier tube has a plurality of anodes as a plurality of pixels, and can output a detection signal according to the amount of light received by each anode. MPPC (Multi-Pixel Photon Counter) is a two-dimensional array of a plurality of pixels, each pixel being an avalanche photodiode that operates in Geiger mode and connected to a quenching resistor. These devices are capable of high-speed, highly sensitive light detection.

The signal processing section 4 inputs the signal output from the photodetector 20. By processing this input signal, the signal processing unit 4 calculates the number n of Cerenkov light detection events on the light receiving surface 21 of the photodetector 20 and the detection position of each Cerenkov light detection event for each interaction event in the radiator 10. Based on (x, y) and the detection time t, the interaction position (x, y, z) in the radiator 10 is estimated by a neural network. Note that the x-axis and y-axis of the x, y, z orthogonal coordinate system are parallel to the light receiving surface 21 of the photodetector 20 and orthogonal to each other. The z-axis is perpendicular to the light receiving surface 21 of the photodetector 20.

The signal processing unit 4 also determines, for each interaction event in the radiator 10, the number n of Cerenkov light detection events on the light receiving surface 21 of the photodetector 20, the detection position (x, y) of each Cerenkov light detection event, and the detection position (x, y) of each Cherenkov light detection event. Based on time t, training data for training the neural network is created. Then, the signal processing unit 4 causes the neural network to learn using this learning data.

The signal processing section 4 is configured by, for example, a computer. The signal processing unit 4 may perform neural network processing using a CPU (Central Processing Unit), but it is preferable to perform processing using a DSP (Digital Signal Processor) or GPU (Graphics Processing Unit) that can perform faster processing. be.

The signal processing section 4 includes a preprocessing section 41 , an estimation section 42 , a learning data creation section 43 , a learning section 44 , and a storage section 45 .

The preprocessing unit 41 determines the detection position (x, y) and detection time t for each Cerenkov light detection event based on the signal output from the photodetector 20. The detection time t can be determined from the pulse timing of the input signal. After that, the preprocessing unit 41 sorts the data including the detection position (x, y) and detection time t of each Cherenkov light detection event by the detection time t. Then, the preprocessing unit 41 determines that data within a time window of a certain width among the sorted data is data of a Cherenkov light detection event caused by a common interaction event.

In this way, the preprocessing unit 41 classifies data including the detection position (x, y) and detection time t of each Cerenkov light detection event for each interaction event. Thereby, the number n of Cerenkov light detection events on the light receiving surface 21, the detection position (x, y) and the detection time t of each Cerenkov light detection event can be determined for each interaction event.

Further, the preprocessing unit 41 discards data of interaction events for which the number n of Cerenkov light detection events is less than the threshold th, and the estimation unit 42 discards data for interaction events for which the number n of Cerenkov light detection events is greater than or equal to the threshold th. Preferably, interaction positions are estimated selectively for action events. This is because, in the case of an interaction event in which the number n of Cerenkov light detection events is less than the threshold th, the estimation unit 42 estimates the interaction position with poor accuracy. Similarly, the learning data creation unit 43 preferably creates learning data for interaction events in which the number n of Cerenkov light detection events is greater than or equal to the threshold th. For example, the threshold value th is 3.

The estimation unit 42 inputs the data preprocessed by the preprocessing unit 41 and calculates the number n of Cerenkov light detection events on the light receiving surface 21 of the photodetector 20 and each Cerenkov light for each interaction event. The interaction position (x, y, z) is estimated by a neural network based on the detection position (x, y) of the detection event and the detection time t. The neural networks used here include, for example, a multi-layer perceptron (MLP), a convolutional neural network (CNN), a recurrent neural network (RNN), and the like.

The estimation unit 42 may include a plurality of neural networks, and may estimate the interaction position using the neural network corresponding to the number n of Cerenkov light detection events. That is, for an interaction event in which the number of Cerenkov light detection events is n, the interaction position may be estimated by a neural network provided corresponding to the number n.

The learning data creation unit 43 generates a plurality of sets of less than n Cerenkov light detection events from among the n Cerenkov light detection events by the photodetector 20 of the Cherenkov detector 3 for each interaction event, The detection position and detection time of each Cerenkov light detection event included in each of these plurality of sets is used as learning data. When there are n Cherenkov photodetection events included in the interaction event and a set of m Cherenkov photodetection events is to be generated from them (m<n), m are selected from the n set. The number of combinations to be extracted is _n C _m .

The learning unit 44 causes the neural network to learn using the learning data created by the learning data creation unit 43. The storage unit 45 stores data before, during, and after processing by the preprocessing unit 41, the estimation unit 42, the learning data creation unit 43, and the learning unit 44, respectively.

FIG. 3 is a diagram illustrating an example of processing by the estimation unit 42 of the signal processing unit 4 when estimating the interaction position. The estimation unit 42 receives data for each interaction event after being preprocessed by the preprocessing unit 41 . That is, the estimation unit 42 inputs data (x, y, t) for each of the n Cerenkov light detection events included in each interaction event. The estimating unit 42 selects data of five Cherenkov light detection events with early detection timing from among the n Cherenkov light detection events included in each interaction event, and calculates the data of the five Cherenkov light detection events ( x, y, t) into the trained neural network. Thereby, the data output from the neural network represents the result of estimating the interaction position (x, y, z) on the radiator 10.

FIG. 4 is a diagram illustrating an example of processing by the learning data creation unit 43 and learning unit 44 of the signal processing unit 4 during learning of the neural network. The learning data creation unit 43 receives data for each interaction event after being preprocessed by the preprocessing unit 41 . That is, the learning data creation unit 43 inputs data (x, y, t) for each of n Cerenkov light detection events included in each interaction event.

The learning data creation unit 43 generates a plurality of sets of five Cerenkov light detection events from among the n Cerenkov light detection events included in each interaction event. At this time, a maximum of _n C ₅ sets can be generated. The learning data creation unit 43 sets the data (x, y, t) of each of the five Cerenkov light detection events included in each group as learning data, and inputs the learning data to the neural network. As a result, the same number of output data (x, y, z) as the number of sets can be obtained from the neural network. The learning unit 44 evaluates the error between the data (x, y, z) output from the neural network and the teacher data (X, Y, Z), and uses the error backpropagation method to reduce the error. Train the neural network by adjusting its parameters.

FIG. 5 is a diagram showing an example of the configuration of a neural network. The neural network shown in this figure is a multilayer perceptron (MLP) with a five-layer structure including an input layer, an output layer and three hidden layers. This MLP inputs data (x, y, t) for each of the five Cerenkov photodetection events and outputs the result of estimating the interaction position (x, y, z). Note that the output data from the neural network may represent a position or may represent the reliability of interaction at each position.

Next, the results of a simulation performed using the Monte Carlo method for the first embodiment will be explained. The radiator 10 was assumed to be made of PbF ₂ and have a rectangular parallelepiped shape of 40×40×10 mm ³ . The light receiving surface 21 of the photodetector 20 was in contact with a 40×40 mm ² surface of the radiator 10. As the photodetector 20, an ideal one with both spatial resolution and temporal resolution is assumed to be infinitesimal. The MLP shown in FIG. 5 was used as a neural network, and data representing the position was output from the MLP.

In the simulation, the interaction of each particle in the radiator is tracked one by one, and by referring to the history, the interaction position in the radiator can be determined, so this can be used as training data (X, Y, Z). can.

For experiments, for example, collimated gamma rays (pencil beam) are used. The position when this pencil beam is incident perpendicularly to the xy plane is used as the interaction position (X, Y). Similarly, when a pencil beam is made perpendicular to the xz plane, (X, Z) is found, and when a pencil beam is made perpendicular to the yz plane, (Y, Z) is found. In this method, since it is not possible to obtain the training data (X, Y, Z) in one go, the neural network for each plane is learned individually and the predictions thereof are integrated. For example, the average of the outputs of the neural networks for (X, Y), (X, Z), and (Y, Z) can be used as the teacher data (X, Y, Z).

The simulation was performed for each of the first comparative example, the second comparative example, and the first example. In the first comparative example, MLP was trained using only data in which the number of Cerenkov light detection events included in interaction events was 5 as training data. In the second comparative example, MLP is performed using data of five Cherenkov light detection events with early detection timings as learning data among data in which the number of Cherenkov light detection events included in the interaction event is 5 or more and 7 or less. Made me learn. In the first example, the learning data created under the conditions described as an example using FIG. I learned MLP.

The number of learning data created in the first comparative example was 521,460. The number of learning data created in the second comparative example was 875,000, which was 1.68 times that in the first comparative example. The number of learning data created in the first example was 4,170,600, which was eight times that of the first comparative example. In this way, compared to the first comparative example and the second comparative example, more learning data can be prepared in the first example.

FIG. 6 is a diagram showing the results of the simulation. This figure shows a histogram of the estimation error of the interaction position in the xy plane and a histogram of the estimation error of the interaction position in the z-axis direction for each of the first comparative example, the second comparative example, and the first example. . The estimation error is the value obtained by subtracting the estimated interaction position from the true interaction position. The horizontal axis represents error, and the vertical axis represents frequency. Each histogram was approximated by a Lorentz function, and the full width at half maximum (FWHM) of the Lorentz function was determined.

In the first comparative example, the FWHM of the estimation error in the xy plane was 0.89 mm, and the FWHM of the estimation error in the z-axis direction was 1.57 mm. In the second comparative example, the FWHM of the estimation error in the xy plane was 0.79 mm, and the FWHM of the estimation error in the z-axis direction was 1.42 mm. In the first example, the FWHM of the estimation error in the xy plane was 0.65 mm, and the FWHM of the estimation error in the z-axis direction was 1.40 mm. Compared with the first comparative example and the second comparative example, in the first example, both the estimation error in the axial direction and the estimation error in the xy plane were small values. In the first example, it is considered that the accuracy of estimating the interaction position is improved by learning the MLP using more training data.

Next, a second embodiment will be described. FIG. 7 is a diagram showing the configuration of the detection system 2 of the second embodiment. The detection system 2 includes a first Cerenkov detector 3A, a second Cerenkov detector 3B, and a signal processing section 5. The Cerenkov detectors 3A and 3B have the same configuration as the Cerenkov detector 3 in the first embodiment. The entrance surfaces 11 (surfaces opposite to the photodetector 20 side, see FIG. 2) of the radiators 10 of the Cerenkov detectors 3A and 3B are opposed to each other.

When an RI source is present in the space between the first Cherenkov detector 3A and the second Cherenkov detector 3B, positrons and electrons emitted from the RI source annihilate, and due to the annihilation, photon pairs ( A pair of gamma rays with an energy of 511 keV) are generated. The two photons that make up the photon pair fly in opposite directions. By detecting one photon by the first Cerenkov detector 3A and the other photon by the second Cerenkov detector 3B, that is, by simultaneously counting the photon pair by the two Cerenkov detectors 3A and 3B. Photon pair generation events (pair annihilation events) can be detected by counting events).

The signal processing unit 5 receives signals output from the photodetectors 20 of each of the Cherenkov detectors 3A and 3B. By processing these input signals, the signal processing unit 5 calculates, for each coincidence event, the number of Cherenkov light detection events on the light receiving surface of the photodetector 20 of each of the Cherenkov detectors 3A and 3B, and the number of Cherenkov light detection events. Based on the detection position and detection time, the photon pair generation position (pair annihilation position) is estimated by a neural network.

Further, the signal processing unit 5 performs a process based on the number of Cherenkov light detection events on the light receiving surface of the photodetector 20 of each of the Cherenkov detectors 3A and 3B and the detection position and detection time of each Cherenkov light detection event for each coincidence event. Create training data for training the neural network. Then, the signal processing unit 5 causes the neural network to learn using this learning data.

The signal processing unit 5 is configured by, for example, a computer. Although the signal processing unit 5 may perform processing using a neural network using a CPU, it is preferable to perform processing using a DSP or GPU that can perform higher-speed processing.

The signal processing section 5 includes a preprocessing section 51 , an estimation section 52 , a learning data creation section 53 , a learning section 54 , and a storage section 55 .

The preprocessing unit 51 performs the same processing as the preprocessing unit 41 in the first embodiment on the signal output from the photodetector 20 of the first Cerenkov detector 3A. The preprocessing unit 51 also performs the same processing as the preprocessing unit 41 in the first embodiment on the signal output from the photodetector 20 of the second Cerenkov detector 3B.

The preprocessing unit 51 also performs the following processing. The preprocessing unit 51 processes interaction event data detected based on the signal output from the photodetector 20 of the first Cerenkov detector 3A and the signal output from the photodetector 20 of the second Cerenkov detector 3B. If the data of the interaction event detected based on the data are within the coincidence time window, it is determined that both data are due to a common photon pair generation event. At this time, it may be determined whether the representative times of both data are within the coincidence time window. The representative time may be the earliest detection time among the detection times included in each interaction event, or may be the average value of the detection times.

The estimation unit 52 inputs the data preprocessed by the preprocessing unit 51 and calculates a Cerenkov light detection event on the light receiving surface of the photodetector 20 of each of the Cerenkov detectors 3A and 3B for each coincidence event. The photon pair generation position (x, y, z) is estimated by a neural network based on the number of , and the detection position and detection time of each Cerenkov photodetection event. The neural network used here is, for example, a multilayer perceptron (MLP), a convolutional neural network (CNN), a recurrent neural network (RNN), or the like.

The learning data creation unit 53 generates, for each coincidence event, a plurality of sets of less than n1 Cherenkov light detection events from among the n1 Cherenkov light detection events by the photodetector 20 of the first Cherenkov detector 3A. Then, the detection position and detection time of each Cerenkov light detection event included in each of these plurality of sets are used as first learning data, and n2 Cerenkov light detections by the photodetector 20 of the second Cerenkov detector 3B are performed. A plurality of sets of less than n2 Cerenkov light detection events are generated from among the events, and the detection position and detection time of each Cerenkov light detection event included in each of the plurality of sets is used as second learning data.

The learning unit 54 causes the neural network to learn using the first learning data and the second learning data created by the learning data creating unit 53. The storage unit 55 stores data before, during, and after processing by the preprocessing unit 51, the estimation unit 52, the learning data creation unit 53, and the learning unit 54, respectively.

FIG. 8 is a diagram illustrating an example of processing by the estimation unit 52 of the signal processing unit 5 when estimating the photon pair generation position. The estimation unit 52 receives data for each coincidence event after being preprocessed by the preprocessing unit 51 . That is, the estimation unit 52 inputs data (x, y, t) for each of n1 Cerenkov light detection events by the first Cerenkov detector 3A for each coincidence event, and inputs n2 data (x, y, t) for each of the n1 Cerenkov light detection events by the second Cerenkov detector 3B. Input data (x, y, t) for each of the Cherenkov photodetection events. The estimation unit 52 selects data of five Cherenkov photodetection events with early detection timings from data of n1 Cherenkov photodetection events by the first Cherenkov detector 3A, and selects data of n2 Cherenkov photodetection events by the second Cherenkov detector 3B. The data of five Cherenkov photodetection events with early detection timings are selected from among the data of each of the Cherenkov photodetection events, and these selected data are input to the trained neural network. Thereby, the data output from the neural network represents the result of estimating the photon pair generation position (x, y, z).

FIG. 9 is a diagram illustrating an example of processing by the learning data creation unit 53 and learning unit 54 of the signal processing unit 5 during learning of the neural network. The learning data creation section 53 receives data for each coincidence event after being preprocessed by the preprocessing section 51 . That is, the learning data creation unit 53 inputs data (x, y, t) for each of n1 Cerenkov light detection events by the first Cerenkov detector 3A for each coincidence event, and also inputs data (x, y, t) for each of n1 Cerenkov light detection events by the first Cerenkov detector 3A Input data (x, y, t) for each of n2 Cerenkov photodetection events by 3B.

The learning data creation unit 53 generates a plurality of sets of five Cherenkov light detection events from among the n1 Cherenkov light detection events by the first Cherenkov detector 3A. At this time, a maximum of _n1 C ₅ sets can be generated. The learning data creation unit 53 sets the data (x, y, t) of each of the five Cerenkov light detection events included in each set as first learning data.

The learning data creation unit 53 generates a plurality of sets of five Cerenkov light detection events from among the n2 Cerenkov light detection events by the second Cherenkov detector 3B. At this time, a maximum of _n2 C ₅ sets can be generated. The learning data creation unit 53 sets the data (x, y, t) of each of the five Cerenkov light detection events included in each set as second learning data.

The learning data creation unit 53 inputs the first learning data and the second learning data to the neural network. The learning data creation unit 53 can input up to _n1 C ₅ × _n2 C ₅ sets of learning data to the neural network. As a result, the same number of output data (x, y, z) as the number of sets can be obtained from the neural network. The learning unit 54 evaluates the error between the data (x, y, z) output from the neural network and the teacher data (X, Y, Z), and uses the error backpropagation method to reduce the error. Train the neural network by adjusting its parameters.

FIG. 10 is a diagram showing an example of the configuration of a neural network. The neural network shown in this figure is a multilayer perceptron (MLP) with a five-layer structure including an input layer, an output layer and three hidden layers. This MLP inputs the data (x, y, t) of each of the five Cherenkov light detection events selected from the n1 Cherenkov light detection events detected by the first Cherenkov detector 3A, and The data (x, y, t) of each of the 5 Cherenkov photodetection events selected from n2 Cherenkov photodetection events detected by the Cherenkov detector 3B are input, and the photon pair generation position (x, y , z). Note that the output data from the neural network may represent a position, or may represent the certainty of the generation of a photon pair at each position.

Next, the results of a simulation performed using the Monte Carlo method for the second embodiment will be explained. The above-described ones were assumed as the first Cherenkov detector 3A and the second Cherenkov detector 3B, respectively. The midpoint of the line segment connecting the center positions of the incident surfaces 11 of the first Cherenkov detector 3A and the second Cherenkov detector 3B is the origin of the xyz orthogonal coordinate system, and the line segment is on the z-axis. It was assumed that each side of the entrance surface 11 was parallel to the x-axis or the y-axis, and the interval between both entrance surfaces 11 was 50 mm. The MLP shown in FIG. 10 was used as a neural network, and data representing the certainty of photon pair generation at each position was output from the MLP.

As photon pair generation positions (teacher data (X, Y, Z)), 125 (=5 x 5 x 5) positions are set at 5 mm intervals in the positive and negative directions of the x-axis, y-axis, and z-axis, centering on the origin. I assumed. Considering the symmetry of three-dimensional space, an RI source is placed in one quadrant (and on the axis of the boundary of the quadrant) to generate photon pairs and collect coincidence information. By using this information, we created coincidence information for other quadrants as well.

The simulation was performed for each of the third comparative example, the fourth comparative example, and the second example. In the third comparative example, MLP was trained using only data in which the numbers n1 and n2 of interaction events included in a common photon pair occurrence event were both 5 as learning data. In the fourth comparative example, among the data in which the numbers n1 and n2 of interaction events included in a common photon pair generation event are both 5 or more and 8 or less, data of five Cherenkov photodetection events with early detection timings are used. This was used as training data to train MLP. The second example was created under the conditions described as an example using FIG. 9 based on data in which the numbers n1 and n2 of interaction events included in a common photon pair generation event are both 5 or more and 8 or less. MLP was trained using the training data.

The number of learning data created in the third comparative example was 43,000. The number of learning data created in the fourth comparative example was 122,125, which was 2.84 times that in the third comparative example. The number of learning data created in the second example was 6,666,375, which was 155 times that in the third comparative example. In this way, compared to the third comparative example and the fourth comparative example, more learning data can be prepared in the second example.

FIG. 11 is a diagram showing the results of the simulation. This figure shows the accuracy of estimating the photon pair generation position in each of the third comparative example, the fourth comparative example, and the second example. In the third comparative example, the estimation accuracy was 91.45%, in the fourth comparative example, the estimation accuracy was 93.80%, and in the second example, the estimation accuracy was 95.11%. Compared with the third comparative example and the fourth comparative example, in the second example, the accuracy of estimating the photon pair generation position was high. In the second example, it is considered that the accuracy of estimating the photon pair generation position was improved by learning the MLP using more learning data.

Note that if there is a bias in the number of learning data depending on the photon pair generation position, it is conceivable that the estimation of the photon pair generation position by a neural network trained using the biased learning data will also be biased. Therefore, we collected coincidence data while checking the number of learning data created for each photon pair generation position between two Cerenkov detectors, and we When the number reaches a certain target value, it is preferable to stop collecting coincidence data for that photon pair generation position. In this case, the time required to collect coincidence data for each photon pair generation position may vary depending on the position.

As described above, according to the present embodiment, in a detection system including a Cerenkov detector and a signal processing unit using a neural network, a large amount of training data used for training the neural network can be created in a short time. The neural network can be easily trained. Training of the neural network can be easily performed by the manufacturer of the detection system, and can also be easily performed by the user of the detection system. For example, if the detection system needs to be calibrated, the user can easily retrain or additionally train the neural network.

DESCRIPTION OF SYMBOLS 1, 2...Detection system, 3...Cherenkov detector, 3A...1st Cherenkov detector, 3B...2nd Cherenkov detector, 10...radiator, 20...photodetector, 4...signal processing unit, 41...preprocessing Section, 42...Estimation unit, 43...Learning data creation unit, 44...Learning unit, 45...Storage unit, 5...Signal processing unit, 51...Preprocessing unit, 52...Estimation unit, 53...Learning data creation unit, 54...Learning section, 55...Storage section.

Claims (4)

A radiator that generates Cherenkov light by interaction with incident particles, and a light receiving surface that detects the Cherenkov light generated by the radiator, and a signal indicating the detection position and detection time of the Cherenkov light on the light receiving surface. a Cerenkov detector including a photodetector that outputs;
A signal output from the photodetector is input, and for each interaction event in the radiator, the number of Cerenkov light detection events on the light receiving surface and the detection position and detection time of each Cerenkov light detection event are determined. an estimation unit that estimates an interaction position in the radiator using a neural network;
For each interaction event, a plurality of sets of less than n Cerenkov light detection events are generated from among the n number of Cerenkov light detection events by the photodetector of the Cherenkov detector, and each of the sets is included in each of the plurality of sets. a learning data creation unit that uses the detection position and the detection time of each Cerenkov light detection event as learning data;
a learning unit that causes the neural network to learn using the learning data created by the learning data creation unit;
A detection system comprising: A radiator that generates Cherenkov light by interaction with incident particles, and a light receiving surface that detects the Cherenkov light generated by the radiator, and a signal indicating the detection position and detection time of the Cherenkov light on the light receiving surface. a first Cerenkov detector and a second Cerenkov detector each including a photodetector that outputs a photodetector;
The signal output from the photodetector of each of the first Cherenkov detector and the second Cherenkov detector is input, and the second Based on the number of Cherenkov light detection events on the light receiving surface of each of the first Cherenkov detector and the second Cherenkov detector, and the detection position and detection time of each Cherenkov light detection event, a photon pair generation position is determined by a neural network. an estimator that estimates;
For each coincidence event, a plurality of sets of less than n1 Cherenkov light detection events are generated from among the n1 Cherenkov light detection events by the photodetector of the first Cherenkov detector, and for each of these plurality of sets, The detection position and the detection time of each included Cherenkov light detection event are used as first learning data, and less than n2 Cherenkov light detection events are detected from among the n2 Cherenkov light detection events by the photodetector of the second Cherenkov detector. a learning data creation unit that generates a plurality of sets of detection events and uses the detection position and the detection time of each Cerenkov light detection event included in each of the plurality of sets as second learning data;
a learning unit that causes the neural network to learn using the first learning data and the second learning data created by the learning data creation unit;
A detection system comprising: A radiator that generates Cherenkov light by interaction with incident particles, and a light receiving surface that detects the Cherenkov light generated by the radiator, and a signal indicating the detection position and detection time of the Cherenkov light on the light receiving surface. a Cerenkov detector including a photodetector that outputs;
A signal output from the photodetector is input, and for each interaction event in the radiator, the number of Cerenkov light detection events on the light receiving surface and the detection position and detection time of each Cerenkov light detection event are determined. an estimation unit that estimates an interaction position in the radiator using a neural network;
A method for training the neural network of a detection system comprising:
For each interaction event, a plurality of sets of less than n Cerenkov light detection events are generated from among the n number of Cerenkov light detection events by the photodetector of the Cherenkov detector, and each of the sets is included in each of the plurality of sets. a learning data creation step in which the detection position and the detection time of each Cerenkov light detection event are used as learning data;
a learning step of causing the neural network to learn using the learning data created in the learning data creation step;
A learning method that prepares you. A radiator that generates Cherenkov light by interaction with incident particles, and a light receiving surface that detects the Cherenkov light generated by the radiator, and a signal indicating the detection position and detection time of the Cherenkov light on the light receiving surface. a first Cerenkov detector and a second Cerenkov detector each including a photodetector that outputs a photodetector;
The signal output from the photodetector of each of the first Cherenkov detector and the second Cherenkov detector is input, and the second Based on the number of Cherenkov light detection events on the light receiving surface of each of the first Cherenkov detector and the second Cherenkov detector, and the detection position and detection time of each Cherenkov light detection event, a photon pair generation position is determined by a neural network. an estimator that estimates;
A method for training the neural network of a detection system comprising:
For each coincidence event, a plurality of sets of less than n1 Cherenkov light detection events are generated from among the n1 Cherenkov light detection events by the photodetector of the first Cherenkov detector, and for each of these plurality of sets, The detection position and the detection time of each included Cherenkov light detection event are used as first learning data, and less than n2 Cherenkov light detection events are detected from among the n2 Cherenkov light detection events by the photodetector of the second Cherenkov detector. a learning data creation step of generating a plurality of sets of detection events and using the detection position and the detection time of each Cerenkov light detection event included in each of the plurality of sets as second learning data;
a learning step of causing the neural network to learn using the first learning data and the second learning data created in the learning data creation step;
A learning method that prepares you.

Images